Communication signal modulation identification method and system based on task incremental learning

Abstract

The invention relates to a communication signal modulation identification method and system based on task incremental learning, belongs to the field of communication modulation identification, and solves the problem of low communication signal modulation identification precision in a dynamic spectrum environment caused by a catastrophic forgetting problem. Comprising the following steps: respectively generating modulation signal data of a plurality of tasks in a dynamic spectrum environment, and taking corresponding modulation modes as sample labels to obtain modulation signal data sets under different tasks; a communication signal modulation identification model is initialized, an experience playback buffer area is created, and the experience playback buffer area is used for storing sample data trained each time and weight parameters of the correspondingly obtained model; performing task incremental learning training on the model by using modulation signal data sets under different tasks and sample data stored in the experience playback buffer area to obtain a trained model; communication signals in a dynamic spectrum environment are obtained in real time and input into the trained model, a corresponding communication signal modulation mode recognition result is obtained, and accurate recognition of the modulation mode is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of communication modulation recognition, and in particular to a communication signal modulation recognition method and system based on task incremental learning. Background Art

[0002] Cognitive wireless communication is a technology that improves wireless communication by using artificial intelligence and machine learning. With the continuous development of artificial intelligence technology, cognitive wireless communication will play an increasingly important role in improving wireless communication performance, enhancing network intelligence, and optimizing spectrum management. It relies on intelligent radio technology, which can sense and respond to changes in the surrounding environment in real time, collect data about signal strength, interference level, network traffic and other related information, and then process and analyze these data through artificial intelligence algorithms to automatically adjust system parameters and decisions to optimize the performance of wireless communication. Specifically, if it is detected that the interference on a certain frequency band is large, the system can automatically switch to another frequency band to avoid interference and improve communication quality; in addition, the cognitive wireless communication system can sense and learn the real-time state of the wireless channel, and dynamically adjust the transmission parameters (such as modulation mode, symbol rate, coding mode, transmission power, etc.) to adapt to the current channel conditions. When the channel quality is poor, the system will use a lower modulation mode and a higher coding rate to ensure the reliability of transmission; when the channel quality is good, the system will use a higher modulation mode and a lower coding rate to achieve higher transmission efficiency. In this environment, wireless communication equipment can dynamically switch between different transmission parameters such as frequency bands, spectrums, and transmission powers, and the communication signal will show dynamic and continuous changes.

[0003] In the field of communication, modulation recognition technology can be used in signal demodulation, decoding and transmission of various communication systems and classification and recognition of communication signals, thereby improving the effectiveness and reliability of communication systems. It is an important signal processing technology for the physical layer of communication. In terms of the network layer, the dynamic spectrum environment is affected by various factors such as internal interference (from other users in the network), external interference (from other communication systems), jammers, channel effects (such as path loss, attenuation, multipath and Doppler effect) and traffic congestion, and wireless networks have various forms of damage in communication. In order to support dynamic spectrum access, users within the network need to perceive the spectrum and classify various signals in the dynamic spectrum environment through modulation recognition. With the rapid development of deep learning technology, modulation recognition technology based on deep learning has also received widespread attention and achieved good recognition performance, but the communication signal data set in a static and unchanging spectrum environment has achieved a higher recognition accuracy. Due to the catastrophic forgetting problem in deep learning, that is, when the deep learning model learns a new task or adapts to a new environment, it may forget the old knowledge learned previously or the relevant information of the old task, causing the model to lose its previous capabilities and unable to achieve effective and reliable modulation recognition performance on a dynamically and continuously changing communication signal set.

[0004] Continual Learning is a machine learning algorithm that has the ability to continuously update and improve models in a dynamic environment. By simulating the way humans learn, the model is allowed to gradually adapt and improve in the process of continuously receiving new data. It can be divided into task incremental learning, class incremental learning, domain incremental learning, etc. When the model faces new communication signals or new channel conditions, it will forget the previously learned modulation recognition knowledge. The forgetting problem limits the performance of the model in a dynamically changing environment and the accuracy of recognition. Summary of the invention

[0005] In view of the above analysis, an embodiment of the present invention aims to provide a communication signal modulation recognition method and system based on task incremental learning, so as to solve the technical problem of low accuracy of communication signal modulation recognition in a dynamic spectrum environment caused by catastrophic forgetting problem.

[0006] The purpose of the present invention is mainly achieved through the following technical solutions:

[0007] The present invention provides a communication signal modulation recognition method based on task incremental learning, comprising the following steps:

[0008] Generate modulation signal data for multiple tasks in a dynamic spectrum environment, and use the corresponding modulation modes as sample labels to obtain modulation signal data sets for different tasks;

[0009] Initialize the communication signal modulation recognition model and create an experience replay buffer, wherein the experience replay buffer is used to store the sample data trained each time and the corresponding weight parameters of the model; perform task incremental learning training on the communication signal modulation recognition model using the modulation signal data sets under the different tasks and the sample data stored in the experience replay buffer to obtain a trained communication signal modulation recognition model;

[0010] The communication signal in the dynamic spectrum environment is acquired in real time and input into the trained communication signal modulation signal recognition model to obtain the corresponding communication signal modulation mode recognition result.

[0011] Furthermore, the task is to identify the modulation of communication signals under the state of changing transmission parameters in a dynamic spectrum environment; wherein the changes in the transmission parameters include dynamically changing transmission power, symbol rate, coding method and modulation method.

[0012] Furthermore, the modulation signal data of each task is adjusted to obtain the modulation signal data under each task, including:

[0013] Corresponding to the modulated signal data under each task, based on a signal-to-noise ratio range of 0 to m dB, the modulated signal data under each signal-to-noise ratio is generated in increments of n dB;

[0014] Based on the symbol rates from p to qkS / s, in increments of MkS / s, generating the modulated signal data at each symbol rate for the modulated signal data at each signal-to-noise ratio;

[0015] Generate an IQ signal of the modulated signal data at each symbol rate, each IQ signal including a plurality of sampling points;

[0016] Gaussian white noise, channel multipath fading, clock offset, frequency offset, sampling rate offset, and Doppler offset channel loss factors are added to each IQ symbol point to simulate a real dynamic spectrum environment, and the modulation signal data for each task is obtained by combining the modulation methods of each signal.

[0017] Among them, m, n, p, q, and M are preset values, and n<m, p<M<q.

[0018] Furthermore, using the modulation signal data sets under the different tasks to perform task incremental learning training on the communication signal modulation recognition model includes:

[0019] The modulated signal data set under different tasks is D = {D1, D2, ... D T}, using the training data set Perform model training, where T is the total number of tasks, 1<T EV <T;

[0020] D CV The sample data in the first task D1 is input into the communication signal modulation recognition model. After the training is completed, the sample data used in the first task and the weight parameters of the obtained model are used as experience data and stored in the experience playback buffer;

[0021] D CV The sample data in the second task D2 and the equivalent amount of data in the first task obtained from the experience replay buffer are input into the communication signal modulation recognition model training. After the training is completed, the sample data used and the weight parameters of the obtained model are stored in the experience replay buffer as experience data;

[0022] During subsequent task training, sample data of tasks that have been trained are obtained from the experience replay buffer according to the amount of sample data of the task, and input together with the sample data of the task into the communication signal modulation recognition model training. After the training is completed, the sample data used this time and the weight parameters of the obtained model are used as experience data and stored in the experience replay buffer; wherein the amount of sample data that has been trained is equal to the amount of sample data of the task; and the sample data that has been trained is evenly extracted from the sample data of each task that has been trained to maintain incremental learning of previous tasks; until the incremental training of all tasks in the training data set is completed.

[0023] Further, using the test data set in the modulated signal data set To verify, in the T EV When training for task T+1, an equal amount of sample data that has been trained is obtained from the experience playback buffer; EV +2 to T-th task verification only uses the sample data in this task;

[0024] After verification using the test data set, the recognition accuracy of the communication signal modulation recognition model is calculated, and when the recognition accuracy reaches a preset threshold, a trained communication signal modulation recognition model is obtained.

[0025] Furthermore, the recognition accuracy A k , calculated as follows:

[0026]

[0027] Among them, a j T is learned for continuous incremental training EV After tasks, the recognition accuracy on the jth task test data.

[0028] Furthermore, the communication signal modulation recognition model is trained using the training data set, and the loss function used is as follows:

[0029]

[0030] in, is the learnable weight parameter for the kth task training in the model, lr k-1 is the learning rate of the k-1th task, D k , D replay They are respectively the sample data during the k-th task training and the experience data obtained from the experience replay buffer.

[0031] Furthermore, the experience replay buffer is created based on a queue mechanism.

[0032] Furthermore, the modulation mode of the communication signal includes BPSK, QPSK, 8PSK, 16QAM, 64QAM, GMSK, MSK, 2FSK, 4FSK, OFDM-BPSK, and OFDM-QPSK.

[0033] The present invention provides a communication signal modulation recognition system based on task incremental learning, comprising:

[0034] The task-driven modulation signal data set generation module is used to generate modulation signal data for multiple tasks in a dynamic spectrum environment, and use the corresponding modulation modes as sample labels to obtain modulation signal data sets under different tasks;

[0035] The task incremental learning network training module is used to initialize the communication signal modulation recognition model and create an experience replay buffer, wherein the experience replay buffer is used to store the sample data trained each time and the corresponding weight parameters of the model; the communication signal modulation recognition model is trained by task incremental learning using the modulation signal data sets under the different tasks and the sample data stored in the experience replay buffer to obtain a trained communication signal modulation recognition model;

[0036] The communication signal modulation real-time identification module is used to obtain the communication signal in a dynamic spectrum environment in real time and input it into the trained communication signal modulation signal identification model to obtain the corresponding communication signal modulation mode identification result.

[0037] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0038] 1. The present invention uses the experience replay buffer to store the experience data of the old tasks, and extracts the experience data of the old tasks at the same time when training the new tasks, and trains them together, thereby reducing the forgetting of the previous task knowledge by the communication signal modulation recognition model when learning the new task; and solves the technical problem of low accuracy of communication signal modulation recognition in a dynamic spectrum environment caused by the catastrophic forgetting problem;

[0039] 2. Through task-driven incremental learning training, the communication signal modulation recognition model of the present invention can adapt to the changes in transmission power, symbol rate, coding mode and modulation mode in a dynamic spectrum environment, thereby improving the recognition accuracy of the communication signal modulation mode;

[0040] 3. By reviewing the empirical data of old tasks during the training process, the model has better generalization ability when facing new or unknown communication signals; it can adapt to the dynamic changes of transmission parameters such as transmission power, symbol rate, coding method and modulation method, and maintain high performance in communication signal modulation method recognition under various conditions;

[0041] 4. The present invention can process and identify the modulation mode of communication signals in real time, meeting the demand for rapid response in real-time communication systems;

[0042] 5. The present invention is flexible and scalable. The method and system design allow easy addition of new modulation modes or adjustment of existing models to adapt to new communication standards.

[0043] 6. The present invention can support both single-user and multi-user environments. The system can effectively process signals in a multi-user environment, and can further distinguish between primary users, secondary users, idle users and illegal users, thus supporting more complex communication scenarios as a basis.

[0044] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like components throughout the drawings.

[0046] Figure 1 This is a flow chart of a communication signal modulation recognition method based on task incremental learning in an embodiment of the present invention;

[0047] Figure 2A schematic diagram of the catastrophic forgetting problem of modulation recognition based on deep learning in an embodiment of the present invention;

[0048] Figure 3 The catastrophic forgetting problem of modulation recognition of communication signals under signal-to-noise ratio changes in an embodiment of the present invention is provided;

[0049] Figure 4 It is a training flow chart of a modulation recognition method based on task incremental learning in an embodiment of the present invention;

[0050] Figure 5 Schematic diagram of a modulation recognition system based on task incremental learning in an embodiment of the present invention; DETAILED DESCRIPTION

[0051] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0052] Deep learning is widely used in wireless communications. The receiver receives the wireless signal from the transmitter and uses a classifier to identify and classify the modulation method. During the communication process, the signal may be continuous or discrete. When the receiver receives a dynamically changing signal set, the unrestricted change of the parameters of the deep neural network model will cause the model to almost completely forget what it has learned before, resulting in catastrophic forgetting, such as Figure 2 and Figure 3 shown. Figure 2 Demonstrated the catastrophic forgetting process that a deep neural network model will face after performing incremental learning on dynamically changing signals; Figure 3 The impact of signal-to-noise ratio changes on the classification and recognition accuracy of deep neural network models.

[0053] In order to meet the challenges of modulation recognition of cognitive communication signals in a dynamic spectrum environment and the challenge of low modulation recognition accuracy of communication signals in a dynamic spectrum environment caused by catastrophic forgetting, the present invention provides a modulation recognition method and system based on task incremental learning to improve the recognition accuracy of the modulation mode of cognitive communication signals in an actual dynamic spectrum environment.

[0054] In order to achieve the above technical objectives, the present invention provides the following technical solutions.

[0055] Embodiment 1:

[0056] A specific embodiment of the present invention discloses a communication signal modulation recognition method based on task incremental learning, such as Figure 1 , Figure 4 As shown, the following steps are included:

[0057] Step S1, generating modulation signal data of multiple tasks respectively in a dynamic spectrum environment, and using the corresponding modulation modes as sample labels to obtain modulation signal data sets under different tasks;

[0058] Step S2, initializing the communication signal modulation recognition model, creating an experience replay buffer, the experience replay buffer being used to store the sample data trained each time and the corresponding weight parameters of the model; performing task incremental learning training on the communication signal modulation recognition model using the modulation signal data sets under the different tasks and the sample data stored in the experience replay buffer, to obtain a trained communication signal modulation recognition model;

[0059] Step S3: acquiring the communication signal in a dynamic spectrum environment in real time and inputting the trained communication signal modulation signal recognition model to obtain the corresponding communication signal modulation mode recognition result.

[0060] Step S1, specifically.

[0061] In a dynamic spectrum environment, the modulation signal dataset is prepared task-driven for communication signal modulation mode recognition training.

[0062] The task is to identify the modulation of communication signals under the state of changing transmission parameters in a dynamic spectrum environment; wherein the changes in the transmission parameters include dynamically changing transmission power, symbol rate, coding method and modulation method.

[0063] The modulation modes of the communication signal include BPSK (Binary Phase Shift Keying), QPSK (Quadrature Phase Shift Keying), 8PSK (8-Phase ShiftKeying), 16QAM (16-Quadrature Amplitude Modulation), 64QAM (64-Quadrature Amplitude Modulation), GMSK (Gaussian Minimum Shift Keying), MSK (Minimum ShiftKeying), 2FSK (2-Frequency Shift Keying), 4FSK (4-Frequency Shift Keying), OFDM-BPSK (Orthogonal Frequency-Division Multiplexing-Binary Phase ... Keying, orthogonal frequency division multiplexing combined with binary phase keying), OFDM-QPSK (Orthogonal Frequency-Division Multiplexing-Quadrature Phase Shift Keying).

[0064] Modulation modes are optional, including but not limited to the above 11 modulation modes; these modulation modes each have their own advantages and disadvantages and are suitable for different communication scenarios and business requirements.

[0065] The modulation signal data of each task is adjusted to obtain the modulation signal data under each task, including:

[0066] Corresponding to the modulated signal data under each task, based on a signal-to-noise ratio range of 0 to m dB, the modulated signal data under each signal-to-noise ratio is generated in increments of n dB;

[0067] Based on the symbol rates from p to qkS / s, in increments of MkS / s, generating the modulated signal data at each symbol rate for the modulated signal data at each signal-to-noise ratio;

[0068] Generate IQ symbols of the modulated signal data at each symbol rate, each IQ symbol including a plurality of symbol points;

[0069] Gaussian white noise, channel multipath fading, clock offset, frequency offset, sampling rate offset, and Doppler offset channel loss factors are added to each IQ symbol point to simulate a real dynamic spectrum environment, and the modulation signal data for each task is obtained by combining the modulation methods of each signal.

[0070] Among them, m, n, p, q, and M are preset values, and n<m, p<M<q.

[0071] Exemplarily, m=20dB, n=2dB.

[0072] (1) For the SNR (Signal-to-Noise Ratio) range of 0 to 20 dB, the SNR range is traversed with an incremental step of 2 dB, starting from 0 dB and gradually increasing to 20 dB, increasing by 2 dB each time, with a total of 11 different SNR levels of 0, 2, 4, 6, 8, ..., 20;

[0073] (2) Generate IQ (In-phase, in-phase; Quadrature, orthogonal) symbols for each modulation type at each signal-to-noise ratio, with each modulation type having 1024 IQ symbol points. The IQ symbol is a complex representation of the modulated signal and is commonly used in digital modulation technology.

[0074] (3) Gaussian white noise, channel multipath fading, clock offset, frequency offset, sampling rate offset, Doppler offset and other channel loss factors are added to each symbol point to generate communication signal data. Each modulation method contains 11,000 frames.

[0075] Gaussian white noise, channel multipath fading, clock offset, frequency offset, sampling rate offset, Doppler offset and other channel loss factors are added to each symbol point to simulate and reproduce various interference and distortion conditions that the signal may encounter in a real communication environment.

[0076] For generating modulated signal data at each signal-to-noise ratio at a dynamically changing symbol rate. Exemplarily, p=100 kS / s, q=500 kS / s, M=100. Based on a symbol rate of 100 to 500 kS / s, the modulated signal data at each symbol rate is generated for the modulated signal data at each signal-to-noise ratio in increments of 100 kS / s.

[0077] In a dynamic spectrum environment, the modulation signal data of multiple tasks are generated respectively, and the corresponding modulation modes are used as sample labels to obtain the modulation signal data sets under different tasks. The modulation signal data for the same task are independent and have the same distribution.

[0078] Exemplarily, the modulation signal data set is divided into a training data set and a test data set according to the 80:20 principle; the training data set and the test data set are used for training and verification of the communication signal modulation recognition model, respectively.

[0079] Take the modulation recognition task in a single-user dynamic spectrum environment and the modulation recognition task in a multi-user dynamic spectrum environment as examples. Simulate the dynamic spectrum environment of a single user, take the signal-to-noise ratio and symbol rate changes of a single user as an example, and take the task setting under the dynamic change of SNR in the range of 0 to 20 dB as an example, as shown in Table 1.

[0080] Table 1: Single-user SNR task examples

[0081]

[0082] Add the training sample data of these three tasks to train the model, then the modulated signal dataset D = {D1, D2, ... D T}, T = 3, select T EV =2,D CV ={D1,D2,} is the training sample data, D EV ={D3} is the test sample data.

[0083] Taking the task setting of modulation recognition under the dynamic change of symbol rate at 100kS / s, 200kS / s, and 300kS / s as an example, the symbol rate can be selected as shown in Table 2.

[0084] Table 2: Single-user symbol rate variation task example

[0085]

[0086]

[0087] To simulate the dynamic spectrum environment under multi-user, taking into account that different users may use different modulation methods when transmitting on a single channel, and only focusing on the dynamic changes of modulation recognition, three tasks are set. The 11 modulation methods and the generated additive Gaussian white noise in the dataset are assigned to each task and divided into four signal categories: primary user, secondary user, idle user and illegal user, as shown in Table 3.

[0088] Table 3: Examples of different communication signal modulation methods for multiple users

[0089]

[0090] The primary user is granted the highest priority in channel access, and the modulation mode changes dynamically from BPSK and OFDM-BPSK to QPSK;

[0091] The secondary user tries to access an idle channel and is granted access when the channel is idle, which improves spectrum utilization. The modulation mode is 2FSK, 4FSK, and then dynamically changes to GMSK and then to MSK;

[0092] Idle users do not transmit signals, and the received signals are Gaussian white noise with no dynamic changes;

[0093] At the same time, the wireless network denies access to illegal users. Illegal users attempt to eavesdrop, affecting communications. After successful eavesdropping, the modulation mode is 8PSK, which dynamically changes to 16QAM, then changes to 64QAM, and dynamically changes to MSK.

[0094] Task 1 is to identify the modulation modes of the communication signals of the primary user, secondary user, idle user and illegal user by classifying BPSK, OFDM-BPSK, 2FSK, 4FSK, Gaussian white noise and 8PSK;

[0095] Task 2 is to classify QPSK, GMSK, Gaussian white noise, and 16QAM to identify the modulation modes of communication signals of primary users, secondary users, idle users, and illegal users after the modulation mode changes dynamically;

[0096] Task 3 is to classify OFDM-QPSK, MSK, Gaussian white noise, and 64QAM after the modulation mode continues to change dynamically to realize the modulation mode identification of the communication signals of the primary user, secondary user, idle user, and illegal user.

[0097] Tasks are independent and have no intersection, and each task is identified by a task ID number.

[0098] The generated continuous task modulation signal data set D = {D1, D2, ... D T}, where T is the total number of tasks, and the data set of the kth task is expressed as in is the modulation signal data of the kth task, k is the task ID, and the modulation data of the same task are independent and have the same distribution. is the corresponding modulation mode label, and the corresponding is a sample in the modulated signal dataset. Modulated data of different tasks have different distributions. k is the number of modulation signal samples of the kth task.

[0099] The continuous task modulation signal data set D={D1,D2,…D T} is divided into training data sets (D CV ) and the test dataset (D EV ).

[0100] The training data set (D CV ): Contains T EV Task, recorded as The modulated signal data of the task in this dataset can be replayed multiple times as needed, and is consistent with the test dataset D EV The tasks in a task may have different degrees of similarity.

[0101] The test data set (D EV ): including TT EV Task, recorded as With D CV Different, D EV The tasks in D are used only once, and the learner will observe examples in sequence and be tested throughout the learning process. The final performance will be calculated from EV The test validation is performed on a retained test dataset extracted from the dataset.

[0102] The function of step S1 is to generate a communication signal data set containing multiple modulation modes and channel conditions based on task drive, which is used to train and verify the communication signal modulation recognition model of step S2 to adapt to modulation mode recognition in a dynamic spectrum environment.

[0103] Step S2, including steps S21-S24, specifically.

[0104] Step S21, initializing a communication signal modulation recognition model.

[0105] The communication signal modulation recognition model is a deep neural network model. Exemplarily, the communication signal modulation recognition model is constructed based on a CNN (Convolutional Neural Network) neural network model.

[0106] Among them, θ is a learnable weight parameter. For θ, the initial task weight can be randomly initialized or the initial weight can be defined; the weight of the subsequent task uses the weight parameter of the previous task as the initial value; For the communication signal modulation recognition model, the modulated signal sample data Mapped to the corresponding modulation mode label is the modulation signal data of the kth task, k is the task ID, the modulation data of the same task are independent and have the same distribution, is the corresponding modulation sample label, and the corresponding is a sample in the modulated signal dataset. Modulated data of different tasks have different distributions.

[0107] For the initial task (the first task), the model weights are randomly initialized; the model weights of subsequent tasks use the learnable weight parameters of the previous task as initial values.

[0108] Step S22: Create an experience playback buffer.

[0109] The experience replay buffer is a data structure, and the experience replay buffer is created based on a queue mechanism.

[0110] You can also create an experience playback buffer based on a circular array, depending on specific business needs.

[0111] The experience replay buffer can be stored locally or in the cloud.

[0112] The sample data used in each model training and the corresponding model weight parameters are stored in the experience replay buffer as experience data. The experience replay buffer should have sufficient capacity to store experience data of multiple tasks to mitigate catastrophic forgetting and ensure that the model retains the knowledge of previous old tasks when learning new tasks.

[0113] The experience replay buffer capacity is D replay , in order to store the experience data of T tasks, give each task m = D replay / T size capacity, store m size experience data of each task, experience data includes sample data of each old task and learned network weights and hyperparameters (learning rate lr, etc.). While learning new tasks, sample data in the experience replay buffer are extracted and input into the communication signal modulation recognition model for training, so as to weaken the impact of catastrophic forgetting.

[0114] If the experience replay buffer is full, new data will overwrite the oldest data based on the queue mechanism.

[0115] Step S23: Use the training data set to train the communication signal modulation recognition model.

[0116] Using the modulation signal data sets under the different tasks to perform task incremental learning training on the communication signal modulation recognition model includes:

[0117] The modulated signal data set under different tasks is D = {D1, D2, ... D T}, using the training data set Perform model training, where T is the total number of tasks, 1<T EV <T;

[0118] D CVThe sample data in the first task D1 is input into the communication signal modulation recognition model. After the training is completed, the sample data used in the first task and the weight parameters of the obtained model are used as experience data and stored in the experience playback buffer;

[0119] D CV The sample data in the second task D2 and the equivalent amount of data in the first task obtained from the experience replay buffer are input into the communication signal modulation recognition model training. After the training is completed, the sample data used and the weight parameters of the obtained model are stored in the experience replay buffer as experience data;

[0120] During subsequent task training, sample data of tasks that have been trained are obtained from the experience replay buffer according to the amount of sample data of the task, and input together with the sample data of the task into the communication signal modulation recognition model training. After the training is completed, the sample data used this time and the weight parameters of the obtained model are used as experience data and stored in the experience replay buffer; wherein the amount of sample data that has been trained is equal to the amount of sample data of the task; and the sample data that has been trained is evenly extracted from the sample data of each task that has been trained to maintain incremental learning of previous tasks; until the incremental training of all tasks in the training data set is completed.

[0121] Step S23 includes steps S231-S233.

[0122] Step S231, initial task learning and experience data storage:

[0123] For the k=1th task, modulation mode recognition is performed, and the sample data of the first task in the training data set is input into the communication signal modulation recognition model Train, where θ is a learnable weight parameter that maps the modulated signal samples to the corresponding modulation mode labels.

[0124] The communication signal modulation recognition model training process includes forward propagation, calculation of loss function, and stochastic gradient descent method to obtain model parameters. The standard empirical risk minimization is used to solve the optimal network weight parameters, as shown below:

[0125]

[0126] And D k The sample data and model parameters in are stored in the experience playback buffer; continue to determine whether to perform new task learning. If the next new task learning is entered, k = k + 1, then enter step S232; otherwise, terminate the task incremental training learning and enter step S233.

[0127] After each task is learned, the experience data is stored. After the current task training is completed, the sample data and model parameters used in the training are stored as experience data in the experience playback buffer; the experience data is represented by the task ID as an identifier.

[0128] Step S232: Task switching and experience playback:

[0129] When entering the k+1 task incremental learning training, a new task T is started k+1 Learning, using D k+1 When training the communication signal modulation recognition model, in addition to using the dataset D of the current task k+1 In addition, the datasets D1 to D2 of the old tasks are extracted from the experience replay buffer. k The empirical data from D1 to D k Get data, the total amount of data and the data set D of the current task k+1 The sample data volume of the old task is the same as the sample data volume of the current task. k+1 The stacked data are used as the input of the communication signal modulation recognition model for the k+1th task.

[0130] Therefore, the communication signal modulation recognition model can maintain a balance between the new task and the old task, reducing the parameters of the new task network model. Compared with the previous network model parameters The deviation between them is used to update the model parameters To minimize the new task T k+1 The loss function of .

[0131] The communication signal modulation recognition model is trained using the training data set, and the loss function used is as follows:

[0132]

[0133] in, is the learnable weight parameter for the kth task training in the model, lr k-1 is the learning rate of the k-1th task, D k , D replay They are respectively the sample data during the k-th task training and the experience data obtained from the experience replay buffer.

[0134] Continue to monitor whether there are new tasks to learn and determine whether the current task data is a training sample set. If k≤T EV, the current task data stream is the training sample data, repeat steps SS231 to S232, ensure sufficient training on each task, and iteratively improve the model as needed, which may include hyperparameter adjustment, data enhancement, etc. Otherwise, the task incremental training learning is completed and enter step S233.

[0135] Step S233: save the communication signal modulation recognition model.

[0136] After the task training is completed, save the communication signal modulation recognition model parameters For testing and verification.

[0137] Task incremental learning is completed if k = T EV +1, then use the test task dataset Verify the model

[0138] Step S24: Use the test data set to verify the communication signal modulation recognition model.

[0139] Use the test data set in the modulation signal data set To verify, in the T EV When training for task T+1, an equal amount of sample data that has been trained is obtained from the experience playback buffer; EV +2 to T-th task verification only uses the sample data in this task;

[0140] After verification using the test data set, the recognition accuracy of the communication signal modulation recognition model is calculated, and when the recognition accuracy reaches a preset threshold, a trained communication signal modulation recognition model is obtained.

[0141] In progress EV When training +1 task, obtain the same amount of sample data in the training task from the experience playback buffer, and verify it by combining the new task data with the old task data;

[0142] Then the subsequent T EV +2 to T-th task verification, only the sample data in the task is used to verify the modulation mode recognition performance of unseen signals, improve the robustness of the model, and enable the model to better generalize to new dynamic spectrum environments.

[0143] The test data set is used to evaluate the performance of the communication signal modulation recognition model. The performance indicators may include: calculation accuracy, loss value, precision and other performance indicators.

[0144] Take recognition accuracy as an example to evaluate model performance.

[0145] The recognition accuracy rate A k , calculated as follows:

[0146]

[0147] Among them, a j T is learned for continuous incremental training EV After tasks, the recognition accuracy on the jth task test data.

[0148] Embodiment 2:

[0149] Another embodiment of the present invention discloses a communication signal modulation identification system based on task incremental learning, thereby realizing a communication signal modulation identification method based on task incremental learning in embodiment 1. The specific implementation of each module refers to the corresponding description in embodiment 1.

[0150] like Figure 5 As shown, the system includes a task-driven modulation signal data set generation module M1, a task incremental learning network training module M2 and a communication signal modulation real-time identification module M3.

[0151] The task-driven modulation signal data set generation module M1 is used to generate modulation signal data for multiple tasks in a dynamic spectrum environment, and obtain modulation signal data sets under different tasks with corresponding modulation modes as sample labels;

[0152] The task incremental learning network training module M2 is used to initialize the communication signal modulation recognition model and create an experience replay buffer, wherein the experience replay buffer is used to store the sample data trained each time and the corresponding weight parameters of the model; the communication signal modulation recognition model is trained by task incremental learning using the modulation signal data sets under the different tasks and the sample data stored in the experience replay buffer to obtain a trained communication signal modulation recognition model;

[0153] The communication signal modulation real-time identification module M3 is used to obtain the communication signal in a dynamic spectrum environment in real time and input it into the trained communication signal modulation signal identification model to obtain the corresponding communication signal modulation mode identification result.

[0154] In summary, a communication signal modulation recognition method and system based on task incremental learning according to an embodiment of the present invention has the following beneficial effects:

[0155] 1. The present invention uses the experience replay buffer to store the experience data of the old tasks, and extracts the experience data of the old tasks at the same time when training the new tasks, and trains them together, thereby reducing the forgetting of the previous task knowledge by the communication signal modulation recognition model when learning the new task; and solves the technical problem of low accuracy of communication signal modulation recognition in a dynamic spectrum environment caused by the catastrophic forgetting problem;

[0156] 2. Through task-driven incremental learning training, the communication signal modulation recognition model of the present invention can adapt to the changes in transmission power, symbol rate, coding mode and modulation mode in a dynamic spectrum environment, thereby improving the recognition accuracy of the communication signal modulation mode;

[0157] 3. By reviewing the empirical data of old tasks during the training process, the model has better generalization ability when facing new or unknown communication signals; it can adapt to the dynamic changes of transmission parameters such as transmission power, symbol rate, coding method and modulation method, and maintain high performance in communication signal modulation method recognition under various conditions;

[0158] 4. The present invention can process and identify the modulation mode of communication signals in real time, meeting the demand for rapid response in real-time communication systems;

[0159] 5. The present invention is flexible and scalable. The method and system design allow easy addition of new modulation modes or adjustment of existing models to adapt to new communication standards.

[0160] 6. The present invention can support both single-user and multi-user environments. The system can effectively process signals in a multi-user environment, and can further distinguish between primary users, secondary users, idle users and illegal users, thus supporting more complex communication scenarios as a basis.

[0161] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A communication signal modulation recognition method based on task incremental learning, characterized in that: The steps include: Generate modulation signal data for multiple tasks in a dynamic spectrum environment, and use the corresponding modulation modes as sample labels to obtain modulation signal data sets for different tasks; Initialize the communication signal modulation recognition model and create an experience replay buffer, wherein the experience replay buffer is used to store the sample data trained each time and the corresponding weight parameters of the model; perform task incremental learning training on the communication signal modulation recognition model using the modulation signal data sets under the different tasks and the sample data stored in the experience replay buffer to obtain a trained communication signal modulation recognition model; The communication signal in the dynamic spectrum environment is acquired in real time and input into the trained communication signal modulation signal recognition model to obtain the corresponding communication signal modulation mode recognition result.

2. The method according to claim 1, characterized in that: The task is to identify the modulation of communication signals under the state of changing transmission parameters in a dynamic spectrum environment; wherein the changes in the transmission parameters include dynamically changing transmission power, symbol rate, coding method and modulation method.

3. The method according to claim 1, characterized in that: The modulation signal data of each task is adjusted to obtain the modulation signal data under each task, including: Corresponding to the modulated signal data under each task, based on a signal-to-noise ratio range of 0 to m dB, the modulated signal data under each signal-to-noise ratio is generated in increments of n dB; Based on the symbol rates from p to qkS / s, in increments of MkS / s, generating the modulated signal data at each symbol rate for the modulated signal data at each signal-to-noise ratio; Generate an IQ signal of the modulated signal data at each symbol rate, each IQ signal including a plurality of sampling points; Gaussian white noise, channel multipath fading, clock offset, frequency offset, sampling rate offset, and Doppler offset channel loss factors are added to each IQ signal to simulate a real dynamic spectrum environment, and the modulation signal data for each task is obtained by combining the modulation methods of each signal. Among them, m, n, p, q, and M are preset values, and n<m, p<M<q.

4. The method according to claim 3, characterized in that: Using the modulation signal data sets under the different tasks to perform task incremental learning training on the communication signal modulation recognition model includes: The modulated signal data set under different tasks is D = {D1, D2, ... D T }, using the training data set Perform model training, where T is the total number of tasks, 1<T EV <T; D CV The sample data in the first task D1 is input into the communication signal modulation recognition model. After the training is completed, the sample data used in the first task and the weight parameters of the obtained model are used as experience data and stored in the experience playback buffer; D CV The sample data in the second task D2 and the equivalent amount of data in the first task obtained from the experience replay buffer are input into the communication signal modulation recognition model training. After the training is completed, the sample data used and the weight parameters of the obtained model are stored in the experience replay buffer as experience data; During subsequent task training, sample data of tasks that have been trained are obtained from the experience replay buffer according to the amount of sample data of the task, and input together with the sample data of the task into the communication signal modulation recognition model training. After the training is completed, the sample data used this time and the weight parameters of the obtained model are used as experience data and stored in the experience replay buffer; wherein the amount of sample data that has been trained is equal to the amount of sample data of the task; and the sample data that has been trained is evenly extracted from the sample data of each task that has been trained to maintain incremental learning of previous tasks; until the incremental training of all tasks in the training data set is completed.

5. The method according to claim 4, characterized in that: Use the test data set in the modulation signal data set To verify, in the T EV +1 task training, obtain the same amount of sample data that has been trained from the experience playback buffer; EV +2 to T-th task verification only uses the sample data in this task; After verification using the test data set, the recognition accuracy of the communication signal modulation recognition model is calculated, and when the recognition accuracy reaches a preset threshold, a trained communication signal modulation recognition model is obtained.

6. The method according to claim 5, characterized in that: The recognition accuracy rate A k , calculated as follows: Among them, a j T is learned for continuous incremental training EV After tasks, the recognition accuracy on the jth task test data.

7. The method according to claim 4, characterized in that: The communication signal modulation recognition model is trained using the training data set, and the loss function used is as follows: in, is the learnable weight parameter for the kth task training in the model, lr k-1 is the learning rate of the k-1th task, D k , D replay They are respectively the sample data during the k-th task training and the experience data obtained from the experience replay buffer.

8. The method according to claim 1, characterized in that: The experience replay buffer is created based on a queue mechanism.

9. The method according to any one of claims 1 to 8, characterized in that: The modulation modes of the communication signal include BPSK, QPSK, 8PSK, 16QAM, 64QAM, GMSK, MSK, 2FSK, 4FSK, OFDM-BPSK, and OFDM-QPSK.

10. A communication signal modulation recognition system based on task incremental learning, characterized in that: include: The task-driven modulation signal data set generation module is used to generate modulation signal data for multiple tasks in a dynamic spectrum environment, and use the corresponding modulation modes as sample labels to obtain modulation signal data sets under different tasks; The task incremental learning network training module is used to initialize the communication signal modulation recognition model and create an experience replay buffer, wherein the experience replay buffer is used to store the sample data trained each time and the corresponding weight parameters of the model; the communication signal modulation recognition model is trained by task incremental learning using the modulation signal data sets under the different tasks and the sample data stored in the experience replay buffer to obtain a trained communication signal modulation recognition model; The communication signal modulation real-time identification module is used to obtain the communication signal in a dynamic spectrum environment in real time and input it into the trained communication signal modulation signal identification model to obtain the corresponding communication signal modulation mode identification result.

Images